The debate is shifting from “Will AI replace traditional models?” to the question “How to effectively combine different approaches?”
At the Bank of Russia's scientific conference “Macro Conference: current issues of macroeconomics” held on February 17 in Innopolis, the central topic was the penetration of artificial intelligence technology into economic analysis and forecasting. What yesterday seemed like the preserve of research laboratories is today an everyday tool for regulators and major banks. Realnoe Vremya's correspondent reviewed the main topics of discussion and found out how artificial intelligence is changing the work of economists, asking Central Bank Vice President Alexei Zabotkin about hot topics related to AI.
Three AI revolutions in macroeconomic analysis
The world's leading economists and managers agree: artificial intelligence is radically changing macroeconomic analysis and forecasting methods. According to analysis by Bloomberg Economics, this transformation is taking place in three main areas:
Expand the data boundaries – extract valuable information from unstructured sources (news, social networks, reports). Speed up the research process – automating routine tasks frees up time to interpret results. Improving econometric models – machine learning methods capture complex non-linear relationships.
Of particular interest are studies by Konstantin Styrin (Harvard PhD, Bank of Russia) on the comparative effectiveness of econometric methods. His research shows that complex methods for averaging dynamic models do not always produce more accurate results than simple approaches—an important reminder for those who tend to absolutize complexity.
Forecasting using ML: 40% higher accuracy
Traditional econometric models—VAR, DSGE—often have difficulty analyzing large data sets and identifying nonlinear relationships, especially during periods of high uncertainty. Machine learning methods come to the rescue.
The Reserve Bank of Australia case (2025) has become a classic example. The economists added 22 text variables extracted from the bank's business relationship reports to the standard Phillips curve. The Lasso model with text features improved wage growth forecasts by nearly 40% compared to the baseline model.
As noted on the sidelines of the conference, similar experiments are being conducted at Russian banks. Data from company reports, analyst notes, and meeting minutes is a huge layer of information that has previously remained outside the scope of quantitative analysis. Today, NLP tools can turn this array into structured prediction tools.
Beware: Real-time LLM is still wrong
However, using large language models requires caution. The San Francisco Fed's ChatMacro Research (2026) warns that ChatGPT's real-time inflation forecasts may be inaccurate and out of date, despite performing well in backtests. The problem is that LLM cannot capture structural change at the moment.
As Dmitry Nazarov (USUE) noted in a recent publication, implementing an LLM in economic analysis requires a new method for assessing quality: validity, reproducibility, stability when moving data, as well as built-in security measures – content filtering, hint checking and the “man in the loop” principle.
Sentiment analysis: from news to social networks
Perhaps the fastest growing application of AI is analyzing unstructured data to gauge economic sentiment.
Central banks are actively deploying these technologies. European uses the Cassandra system to analyze the nuances of financial news and uses gradient boosting to identify early signs of stress in the banking sector. Bank for International Settlements researchers are combining recurrent neural networks with LLM to forecast currency crises two months in advance.
Hybrid models: synthesis as a strategy
The key conclusion that participants in the Innopolis discussion reached was that the future does not belong to “pure” AI, but to hybrid approaches that combine the predictive power of machine learning with the interpretive power of structural models.
The main advantage of structural models (SVAR, DSGE) is their interpretability. They allow us to understand cause and effect relationships, obtain impulse responses, and evaluate the consequences of shocks. AI models often operate like a black box.
A study comparing TVP-SVAR with ML methods on energy market data shows that the improved econometric model and better ML algorithm achieve statistically equal forecasting accuracy. However, only a structural approach allows us to obtain the interpretable results needed to understand the transmission mechanisms of shocks.
Experiment of the Bank of Japan
Of particular interest is the Bank of Japan (2025) experiment, in which LLM was used to create agent-based models (ABMs). In the simulation, LLM plays the roles of consumer and business. Their behavior in response to changes in prices and wages is consistent with classical Keynesian theory – rising real wages lead to increased consumption. This opens up the way to create “synthetic microdata” to test scenarios in environments where real data is scarce.
“Trajectory is more important, not personal decisions”:
After completing the official program, we were able to ask Vice President of the Bank of Russia Alexey Zabotkin some questions that came directly from the topics discussed at the conference.
– San Francisco Fed study shows LLM is wrong in forecasting real-time inflation. How do Central Banks address the AI “black box” problem and maintain the interpretability of its models?
“We are well aware of the limitations when using large language models. For us, it is not just the speed of receiving a signal, but also its interpretability and reliability. Therefore, AI in our work never acts as an independent “advisor” making decisions. We use it as a tool for the main processing of large data sets. For example, when analyzing the labor market – where we lack the immediacy of key statistics formula – we can use machine learning models to process streaming data from job sites. But the results are just raw data, hypotheses. Next, we “open the black box”: we check what characteristics the model draws conclusions from, whether it contradicts economic logic, and whether the anomaly is related to seasonality or technical errors.
We also use ensembles of models, compare results from different algorithms, and maintain human control. The interpretability of structural models acts as a filter: it allows us to filter out the noise that AI generates and leave only information that has a clear business case. A decision maker always sees not just a number from a neural network but also a chain of logical structures leading up to this number. Only this combined approach gives us confidence in the conclusions.
— If the neural network is not yet capable of real-time forecasting, could it be a data quality issue? Do you see inflation in the Wildberries data earlier than in the Rosstat report?
— Rosstat data comes to us with the same operating frequency as any other data: weekly data – once a week, complete data for the month – approximately on the 10th working day. And this is a sufficient level of efficiency for monetary policy purposes. More than enough, considering the lag with which monetary policy affects the economy. The narrative that there are some sources that offer more efficient solutions because someone gets price data two days faster is a mistake. The quality of monetary policy decisions is determined to a much greater extent by the ability to consider all indicators as a whole and not just the price level. We are completely satisfied with the performance of the pricing data. But where we really lack the efficiency and detail of official data is in the labor market sector. This is where we rely more on performance indicators from other sources. And I'll say it right away: it's not just at HeadHunter.
HeadHunter data has recently become so popular that it must be considered very carefully. A job seeker can post five resumes there for different majors. Therefore, the number of resumes on HeadHunter is not a good quantitative measure of what is happening with supply in the labor market. We need to correlate this with other metrics. In terms of job vacancies, HeadHunter is a much better source.
Future context and capabilities of Russia
The Research and Forecasting Department of the Bank of Russia, under the leadership of Alexander Morozov, is actively researching the application of ML methods in macroeconomic forecasting.
At the conference, human resources issues received special attention. Andrey Afonin, Director of the Russian Bank University, presented the top 5 skills that determine the success of a specialist (according to the World Economic Forum):
Analytical and creative thinking. Ability to interact with artificial intelligence. Working with big data. Technology knowledge. Network security.
Particular emphasis is placed on the importance of critical thinking to interact effectively with AI. Models can make mistakes and it is up to them to recognize and correct these errors. As Zabotkin himself emphasized on the Central Bank's official Telegram channel, “inflation is still above the target, not to mention inflation expectations have not fallen much.” In such conditions, relying solely on algorithms is an unacceptable luxury.
Conclude
The scientific conference in Innopolis clearly demonstrated: the discussion is shifting from the question “will AI replace traditional models?” to the question “how to effectively combine different approaches?”
Key findings:
AI does not replace but complements. The predictive power of ML combined with the interpretive power of structural models creates a synergistic effect. Context matters. The success of AI depends on data quality, problem definition, and human oversight. LLM requires careful verification, especially in real time. New data sources – new opportunities. Alternative data from markets, job search and trading platforms provides a real-time picture of economic activity, complementing official statistics. Human resources is everything. The economist of the future must master both econometrics and AI methods, and most importantly, be able to think critically and ask the right questions. As Zabotkin correctly notes, “sharp cuts in base rates could accelerate inflation significantly, and high inflation is much worse for businesses over the long term than high rates.” Such decisions require a balanced, human approach, not automated algorithms.
For Tatarstan, where Innopolis is located – one of Russia's key AI development centers – these topics are of particular importance. The Republic could become a testing ground for developing hybrid approaches to economic analysis, combining the capabilities of Innopolis University, the analytical capabilities of major banks and the expertise of the Bank of Russia.
